LogXide

In the historical sink-verified benchmark runs documented in this repository, LogXide measured several-fold faster than stdlib logging: roughly 5–11× on file logging, scenario- and machine-dependent. That is a record of those runs, not a guarantee for every workload. Powered by Rust.

Familiar stdlib-style API — for common patterns, change one import and getLogger, format strings, and handlers work as expected. It's a near-drop-in, not a strict one: some advanced stdlib behaviors differ (flush now drains/waits, LogRecord/Logger can't be subclassed, custom Formatter subclasses fall back to a slower path). See compatibility.

# Before
import logging

# After
from logxide import logging

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger('myapp')
logger.info('Hello, world!')  # Same code. Historical benchmark results: see docs/benchmarks.md.

Installation

pip install logxide

# With Sentry integration
pip install logxide[sentry]

Performance

LogXide is performance-first: its native Rust handlers dispatch on the GIL-released fast path, formatting and writing without materializing a Python LogRecord. As of 0.2.0 the text-sink wrappers (FileHandler, StreamHandler, RotatingFileHandler) emit through that native Rust path by default. A handler switches to the Python path for a custom Formatter subclass, {/$-style format strings, format features the native formatter doesn't replicate (precision like %.2f, escaped %%, or %(asctime)s with stdlib's default milliseconds), or a handler-level Python filter; a handler constructed with non-default file options (mode, encoding, delay, errors) or an arbitrary stream object skips the Rust writer entirely and runs the pure stdlib path. Behavior stays stdlib-correct on all of these paths, only speed changes. See compatibility.

Benchmarks

Two sink-verified benchmarks, both re-run this session on macOS M4 Max, release build, across Python 3.12.11 and 3.14.2. Sink-verified durable throughput means records the sink actually confirmed after flush(), not records merely enqueued. Numbers are machine-specific and rounded to ranges; baselines are noisy run-to-run (roughly ±40%), so treat the ranges as the signal, not any single figure. CPython 3.12 and 3.14 come out at parity once the test environments match, so the ranges below apply to both.

Benchmark A — native FileHandler vs stdlib. benchmark/perf_vs_stdlib.py, LogXide and stdlib each measured in isolation. FileHandler is synchronous, so these are durable (no async drops). Rounded speedup vs stdlib, comparable on Python 3.12 and 3.14:

Scenario	Speedup vs stdlib
simple	~7–9×
structured	~7–9×
%-args	~5–6×

Benchmark B — durable cross-library sink. benchmark/basic_handlers_benchmark.py, each library in its own subprocess, sink-verified 20,200 / 20,200. Rounded speedup vs stdlib, comparable across Python 3.12 and 3.14:

Sink	Speedup vs stdlib
FILE	~6–11×
ROTATING	~8–14×
STREAM	~5× (async, see note)

STREAM is asynchronous. It reaches ~5× when its queue fully drains, but under a sustained max-rate burst the bounded queue can drop records (one loaded run delivered ~14,420 / 20,200; an idle machine delivered 20,200 / 20,200). Treat STREAM as fast best-effort delivery: call flush() and check get_metrics() to confirm what landed, rather than as a guaranteed durable multiplier.

Async HTTP delivery is accounted honestly on both versions: http_block lands 20,000 / 20,000 (durable), while http_drop_newest delivers ~260 / 20,000 and drops the rest, with emitted == sink_acknowledged + queue_dropped + delivery_failed holding throughout.

A prior draft reported a "Python 3.14 regression" (roughly half the file-path speedup on 3.14). That was a measurement artifact, not a real regression: the 3.14 test environment had sentry-sdk installed while the 3.12 one did not, and importing it pulled in a formatter-less NullHandler that forced process-global caller-frame collection on every log (a ~20% tax that only hit the 3.14 runs). This is fixed in 0.2.1; environment-matched, the two versions are at parity.

For full per-handler p50/p99 latency, cross-library detail, and async accounting, see docs/benchmarks.md.

Works With

LogXide intercepts stdlib logging — most libraries work without changes.

Framework / Library	Status	Notes
Flask	✅	app.logger automatically intercepted
Django	✅	LOGGING dictConfig supported
FastAPI / Uvicorn	✅	All uvicorn loggers intercepted
SQLAlchemy	✅	SQL query logging via echo=True
requests / httpx	✅	HTTP connection logs captured
boto3 / botocore	✅	AWS SDK logs captured
Sentry	✅	Native integration — auto-detects an already-configured SDK
Celery	⚠️	Requires setup_logging signal (guide)
pytest	⚠️	Use caplog_logxide instead of caplog

Full compatibility guide for 20+ libraries →

Statuses above describe the shipped releases. Lifecycle corrections (propagation, handler removal, repeat-interception edge cases) are in flight on a separate work stream and are not yet re-verified, so treat an edge-case interception claim as pending until that work lands; the detailed guide notes per-library caveats.

Built-in Sentry Integration

No extra handlers. No configuration. Just works.

import sentry_sdk
sentry_sdk.init(dsn="your-dsn")

from logxide import logging

logger = logging.getLogger(__name__)
logger.error("This is automatically sent to Sentry")

• Auto-detects a configured Sentry SDK (a call to sentry_sdk.init() must have run first)
• WARNING+ sent as events, INFO as breadcrumbs
• Full stack traces and custom context

An installed-but-unconfigured Sentry SDK does not attach a handler, and (as of 0.2.1) importing it no longer forces process-global caller-frame collection onto unrelated handlers.

Native OpenTelemetry Support

Ship logs to any OTLP-compatible backend with zero dependencies:

from logxide import OTLPHandler

handler = OTLPHandler(
    url="http://localhost:4318/v1/logs",
    service_name="my-service"
)

Quick Start

from logxide import logging

# Basic setup — same API as stdlib
logging.basicConfig(
    level=logging.INFO,
    format='%(asctime)s - %(name)s - %(levelname)s - %(message)s'
)

logger = logging.getLogger('myapp')
logger.info('Hello from LogXide!')
logger.warning('This works exactly like stdlib logging')

Custom fields with extra

logger.info("User logged in", extra={
    "user_id": 12345,
    "ip": "192.168.1.1",
    "metadata": {"browser": "Chrome", "version": 120}
})

HTTP log shipping

from logxide import HTTPHandler

handler = HTTPHandler(
    url="https://logs.example.com",
    global_context={"app": "myapp", "env": "production"},
    transform_callback=lambda records: {
        "logs": [{"msg": r["msg"], "level": r["levelname"]} for r in records]
    }
)

What's Different from stdlib

LogXide reimplements Python's logging in Rust for speed. The API is the same, but some advanced stdlib patterns aren't supported:

Feature	Status
getLogger, info, debug, warning, error, critical	✅ Same API
basicConfig, format strings, levels, filters	✅ Same API
FileHandler, StreamHandler, RotatingFileHandler	✅ Rust-native
HTTPHandler, OTLPHandler	✅ Rust-native, high throughput
Custom Python handlers via addHandler()	⚠️ Accepted; runs once on the Python side (no fast-path GIL release)
Subclassing LogRecord or Logger	❌ Rust types, not subclassable
pytest caplog fixture	⚠️ Use caplog_logxide instead

Instead of subclassing LogRecord, use extra={} for custom fields, global_context for metadata, or transform_callback for output transformation.

Compatibility

• Python: 3.12, 3.13, 3.14 (fully tested and supported)
• Python 3.15: Not yet supported — blocked by an upstream pyo3 ↔ Python 3.15-alpha ABI mismatch (the compiled extension references a CPython internal symbol _PyType_FromSlots that current 3.15 alpha builds do not export). Tracking for re-enablement once pyo3 ships a 3.15-compatible release.
• Platforms: macOS, Linux, Windows
• Dependencies: None (Rust compiled into native extension)

Documentation

• Usage Guide — Complete API guide
• Integration Guide — Flask, Django, FastAPI
• Third-Party Compatibility — 20+ libraries
• Performance Benchmarks — Detailed analysis
• Architecture — Technical design
• API Reference — Full reference

Contributing

git clone https://github.com/Indosaram/logxide
cd logxide
pip install -e '.[dev]'   # includes pytest-timeout, required by the pytest config
maturin develop
pytest tests/

The Rust toolchain is pinned to 1.88.0 in rust-toolchain.toml; rustup installs it automatically. This is the version the project builds with, not a verified minimum supported version.

See development guide for details.

License

MIT License — see LICENSE for details.